GPU economics: how much money does one AI server actually make?

"Data centers print money" is the kind of sentence that should make you suspicious. So let's do what marketing pages rarely do: walk through the actual unit economics of one AI server, line by line, including the parts that eat the profit. The numbers below are illustrative industry ranges, not Tervolt projections — the point is the structure, not the decimals.

The revenue side: renting out compute

An AI server's product is GPU-hours. Market rates for a modern data-center GPU have ranged roughly between $1 and $4 per GPU-hour depending on the chip generation, contract length, and how scarce capacity is. An 8-GPU server rented at, say, $2 per GPU-hour would gross about $16 per hour — around $140,000 per year if it ran fully rented, every hour of the year.

That "if" is the entire business.

Deduction 1: Utilization

No facility rents 100% of its capacity 100% of the time. Contracts have gaps, hardware needs maintenance windows, and demand fluctuates. Long-term contracted capacity might run at high utilization; spot-market capacity swings widely. At 70% utilization, our illustrative server's $140,000 becomes roughly $98,000. Utilization is the single most important operational variable — and it is not fully in the operator's control.

Deduction 2: Power and cooling

AI servers are energy-dense. An 8-GPU server under load draws several kilowatts around the clock, and cooling adds a substantial overhead on top (measured by PUE — power usage effectiveness). Depending on electricity prices — which vary enormously between, say, Scandinavia and Germany — power and cooling can consume anywhere from a modest slice to a painful share of revenue. This is why site selection is a return driver, not a real-estate detail.

Deduction 3: Depreciation — the silent giant

Here is the number most pitches skip: the server itself loses value fast. GPU generations turn over roughly every 18–24 months, and rental rates for older chips fall as newer ones arrive. Most operators depreciate AI hardware over three to five years. On a server costing $250,000+, straight-line depreciation alone can be $50,000–$80,000 per year — often the largest single cost line, bigger than power.

Deduction 4: Everything else

Facility staff, networking, security, insurance, maintenance and replacement parts, financing costs, and platform overhead. Individually small, collectively real.

What's left — and why it's the return

After utilization, energy, depreciation, and operations, the margin that remains is what can be distributed to whoever financed the hardware. Under good conditions — strong utilization, sensible power prices, disciplined refresh planning — the economics support meaningful annual returns on invested capital. Under bad conditions — a demand dip, an energy spike, a faster-than-planned generation shift — the same structure produces low, zero, or negative returns.

This is exactly why serious offerings state target ranges rather than fixed rates, and why any range like 5–15% per year has to be read as a scenario, not a promise. The upper end assumes things going well; the lower end assumes friction; and below the range sits the real possibility of loss, including total loss if a project fails outright.

The three questions this teaches you to ask

1. What utilization does the model assume? If a pitch assumes near-perfect utilization forever, it's a red flag.

2. How is depreciation handled? Ask what refresh cycle the model assumes and who bears the cost of replacing aging GPUs.

3. Where does the power come from and at what price? Cheap, stable (ideally renewable) power is one of the strongest structural advantages a facility can have.

An operator who answers these three questions specifically and in writing is treating you like an investor. One who answers with adjectives is treating you like a mark.

The takeaway

AI compute is a genuinely strong business — demand is extraordinary and revenue per megawatt beats traditional hosting several times over. But it is an operating business with real costs and real volatility, not a money printer. Understanding the unit economics is the best protection an investor has: it turns "trust me" into arithmetic you can interrogate.

Nothing in this article is investment advice. All figures are illustrative industry ranges, not projections or promises. Capital at risk, including total loss.